1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #define DEBUG_SUBSYSTEM S_RPC
39 #include <liblustre.h>
41 #include <obd_support.h>
42 #include <obd_class.h>
43 #include <lustre_net.h>
44 #include <lu_object.h>
45 #include <lnet/types.h>
46 #include "ptlrpc_internal.h"
48 /* The following are visible and mutable through /sys/module/ptlrpc */
49 int test_req_buffer_pressure = 0;
50 CFS_MODULE_PARM(test_req_buffer_pressure, "i", int, 0444,
51 "set non-zero to put pressure on request buffer pools");
52 unsigned int at_min = 0;
53 CFS_MODULE_PARM(at_min, "i", int, 0644,
54 "Adaptive timeout minimum (sec)");
55 unsigned int at_max = 600;
56 EXPORT_SYMBOL(at_max);
57 CFS_MODULE_PARM(at_max, "i", int, 0644,
58 "Adaptive timeout maximum (sec)");
59 unsigned int at_history = 600;
60 CFS_MODULE_PARM(at_history, "i", int, 0644,
61 "Adaptive timeouts remember the slowest event that took place "
62 "within this period (sec)");
63 static int at_early_margin = 5;
64 CFS_MODULE_PARM(at_early_margin, "i", int, 0644,
65 "How soon before an RPC deadline to send an early reply");
66 static int at_extra = 30;
67 CFS_MODULE_PARM(at_extra, "i", int, 0644,
68 "How much extra time to give with each early reply");
72 static int ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc);
74 static CFS_LIST_HEAD(ptlrpc_all_services);
75 spinlock_t ptlrpc_all_services_lock;
78 ptlrpc_alloc_request_buffer (int size)
82 if (size > SVC_BUF_VMALLOC_THRESHOLD)
83 OBD_VMALLOC(ptr, size);
91 ptlrpc_free_request_buffer (char *ptr, int size)
93 if (size > SVC_BUF_VMALLOC_THRESHOLD)
99 struct ptlrpc_request_buffer_desc *
100 ptlrpc_alloc_rqbd (struct ptlrpc_service *svc)
102 struct ptlrpc_request_buffer_desc *rqbd;
108 rqbd->rqbd_service = svc;
109 rqbd->rqbd_refcount = 0;
110 rqbd->rqbd_cbid.cbid_fn = request_in_callback;
111 rqbd->rqbd_cbid.cbid_arg = rqbd;
112 CFS_INIT_LIST_HEAD(&rqbd->rqbd_reqs);
113 rqbd->rqbd_buffer = ptlrpc_alloc_request_buffer(svc->srv_buf_size);
115 if (rqbd->rqbd_buffer == NULL) {
120 spin_lock(&svc->srv_lock);
121 list_add(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
123 spin_unlock(&svc->srv_lock);
129 ptlrpc_free_rqbd (struct ptlrpc_request_buffer_desc *rqbd)
131 struct ptlrpc_service *svc = rqbd->rqbd_service;
133 LASSERT (rqbd->rqbd_refcount == 0);
134 LASSERT (list_empty(&rqbd->rqbd_reqs));
136 spin_lock(&svc->srv_lock);
137 list_del(&rqbd->rqbd_list);
139 spin_unlock(&svc->srv_lock);
141 ptlrpc_free_request_buffer (rqbd->rqbd_buffer, svc->srv_buf_size);
146 ptlrpc_grow_req_bufs(struct ptlrpc_service *svc)
148 struct ptlrpc_request_buffer_desc *rqbd;
151 CDEBUG(D_RPCTRACE, "%s: allocate %d new %d-byte reqbufs (%d/%d left)\n",
152 svc->srv_name, svc->srv_nbuf_per_group, svc->srv_buf_size,
153 svc->srv_nrqbd_receiving, svc->srv_nbufs);
154 for (i = 0; i < svc->srv_nbuf_per_group; i++) {
155 rqbd = ptlrpc_alloc_rqbd(svc);
158 CERROR ("%s: Can't allocate request buffer\n",
163 if (ptlrpc_server_post_idle_rqbds(svc) < 0)
171 ptlrpc_save_lock (struct ptlrpc_request *req,
172 struct lustre_handle *lock, int mode, int no_ack)
174 struct ptlrpc_reply_state *rs = req->rq_reply_state;
178 LASSERT(rs->rs_nlocks < RS_MAX_LOCKS);
180 idx = rs->rs_nlocks++;
181 rs->rs_locks[idx] = *lock;
182 rs->rs_modes[idx] = mode;
183 rs->rs_difficult = 1;
184 rs->rs_no_ack = !!no_ack;
188 ptlrpc_schedule_difficult_reply (struct ptlrpc_reply_state *rs)
190 struct ptlrpc_service *svc = rs->rs_service;
193 LASSERT_SPIN_LOCKED(&svc->srv_lock);
194 LASSERT (rs->rs_difficult);
195 rs->rs_scheduled_ever = 1; /* flag any notification attempt */
197 if (rs->rs_scheduled) { /* being set up or already notified */
202 rs->rs_scheduled = 1;
203 list_del (&rs->rs_list);
204 list_add (&rs->rs_list, &svc->srv_reply_queue);
205 cfs_waitq_signal (&svc->srv_waitq);
210 ptlrpc_commit_replies (struct obd_device *obd)
212 struct list_head *tmp;
213 struct list_head *nxt;
216 /* Find any replies that have been committed and get their service
217 * to attend to complete them. */
219 /* CAVEAT EMPTOR: spinlock ordering!!! */
220 spin_lock(&obd->obd_uncommitted_replies_lock);
222 list_for_each_safe (tmp, nxt, &obd->obd_uncommitted_replies) {
223 struct ptlrpc_reply_state *rs =
224 list_entry(tmp, struct ptlrpc_reply_state, rs_obd_list);
226 LASSERT (rs->rs_difficult);
228 if (rs->rs_transno <= obd->obd_last_committed) {
229 struct ptlrpc_service *svc = rs->rs_service;
231 spin_lock (&svc->srv_lock);
232 list_del_init (&rs->rs_obd_list);
233 ptlrpc_schedule_difficult_reply (rs);
234 spin_unlock (&svc->srv_lock);
238 spin_unlock(&obd->obd_uncommitted_replies_lock);
243 ptlrpc_server_post_idle_rqbds (struct ptlrpc_service *svc)
245 struct ptlrpc_request_buffer_desc *rqbd;
250 spin_lock(&svc->srv_lock);
252 if (list_empty (&svc->srv_idle_rqbds)) {
253 spin_unlock(&svc->srv_lock);
257 rqbd = list_entry(svc->srv_idle_rqbds.next,
258 struct ptlrpc_request_buffer_desc,
260 list_del (&rqbd->rqbd_list);
262 /* assume we will post successfully */
263 svc->srv_nrqbd_receiving++;
264 list_add (&rqbd->rqbd_list, &svc->srv_active_rqbds);
266 spin_unlock(&svc->srv_lock);
268 rc = ptlrpc_register_rqbd(rqbd);
275 spin_lock(&svc->srv_lock);
277 svc->srv_nrqbd_receiving--;
278 list_del(&rqbd->rqbd_list);
279 list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
281 /* Don't complain if no request buffers are posted right now; LNET
282 * won't drop requests because we set the portal lazy! */
284 spin_unlock(&svc->srv_lock);
289 struct ptlrpc_service *ptlrpc_init_svc_conf(struct ptlrpc_service_conf *c,
290 svc_handler_t h, char *name,
291 struct proc_dir_entry *proc_entry,
292 svcreq_printfn_t prntfn,
295 return ptlrpc_init_svc(c->psc_nbufs, c->psc_bufsize,
296 c->psc_max_req_size, c->psc_max_reply_size,
297 c->psc_req_portal, c->psc_rep_portal,
298 c->psc_watchdog_factor,
300 prntfn, c->psc_min_threads, c->psc_max_threads,
301 threadname, c->psc_ctx_tags, NULL);
303 EXPORT_SYMBOL(ptlrpc_init_svc_conf);
305 static void ptlrpc_at_timer(unsigned long castmeharder)
307 struct ptlrpc_service *svc = (struct ptlrpc_service *)castmeharder;
308 svc->srv_at_check = 1;
309 svc->srv_at_checktime = cfs_time_current();
310 cfs_waitq_signal(&svc->srv_waitq);
313 /* @threadname should be 11 characters or less - 3 will be added on */
314 struct ptlrpc_service *
315 ptlrpc_init_svc(int nbufs, int bufsize, int max_req_size, int max_reply_size,
316 int req_portal, int rep_portal, int watchdog_factor,
317 svc_handler_t handler, char *name,
318 cfs_proc_dir_entry_t *proc_entry,
319 svcreq_printfn_t svcreq_printfn,
320 int min_threads, int max_threads,
321 char *threadname, __u32 ctx_tags,
322 svc_hpreq_handler_t hp_handler)
325 struct ptlrpc_service *service;
329 LASSERT (bufsize >= max_req_size + SPTLRPC_MAX_PAYLOAD);
330 LASSERT (ctx_tags != 0);
332 OBD_ALLOC_PTR(service);
336 /* First initialise enough for early teardown */
338 service->srv_name = name;
339 spin_lock_init(&service->srv_lock);
340 CFS_INIT_LIST_HEAD(&service->srv_threads);
341 cfs_waitq_init(&service->srv_waitq);
343 service->srv_nbuf_per_group = test_req_buffer_pressure ? 1 : nbufs;
344 service->srv_max_req_size = max_req_size + SPTLRPC_MAX_PAYLOAD;
345 service->srv_buf_size = bufsize;
346 service->srv_rep_portal = rep_portal;
347 service->srv_req_portal = req_portal;
348 service->srv_watchdog_factor = watchdog_factor;
349 service->srv_handler = handler;
350 service->srv_request_history_print_fn = svcreq_printfn;
351 service->srv_request_seq = 1; /* valid seq #s start at 1 */
352 service->srv_request_max_cull_seq = 0;
353 service->srv_threads_min = min_threads;
354 service->srv_threads_max = max_threads;
355 service->srv_thread_name = threadname;
356 service->srv_ctx_tags = ctx_tags;
357 service->srv_hpreq_handler = hp_handler;
358 service->srv_hpreq_ratio = PTLRPC_SVC_HP_RATIO;
359 service->srv_hpreq_count = 0;
360 service->srv_n_hpreq = 0;
362 rc = LNetSetLazyPortal(service->srv_req_portal);
365 CFS_INIT_LIST_HEAD(&service->srv_request_queue);
366 CFS_INIT_LIST_HEAD(&service->srv_request_hpq);
367 CFS_INIT_LIST_HEAD(&service->srv_idle_rqbds);
368 CFS_INIT_LIST_HEAD(&service->srv_active_rqbds);
369 CFS_INIT_LIST_HEAD(&service->srv_history_rqbds);
370 CFS_INIT_LIST_HEAD(&service->srv_request_history);
371 CFS_INIT_LIST_HEAD(&service->srv_active_replies);
372 CFS_INIT_LIST_HEAD(&service->srv_reply_queue);
373 CFS_INIT_LIST_HEAD(&service->srv_free_rs_list);
374 cfs_waitq_init(&service->srv_free_rs_waitq);
376 spin_lock_init(&service->srv_at_lock);
377 CFS_INIT_LIST_HEAD(&service->srv_req_in_queue);
378 CFS_INIT_LIST_HEAD(&service->srv_at_list);
379 cfs_timer_init(&service->srv_at_timer, ptlrpc_at_timer, service);
380 /* At SOW, service time should be quick; 10s seems generous. If client
381 timeout is less than this, we'll be sending an early reply. */
382 at_init(&service->srv_at_estimate, 10, 0);
384 spin_lock (&ptlrpc_all_services_lock);
385 list_add (&service->srv_list, &ptlrpc_all_services);
386 spin_unlock (&ptlrpc_all_services_lock);
388 /* Now allocate the request buffers */
389 rc = ptlrpc_grow_req_bufs(service);
390 /* We shouldn't be under memory pressure at startup, so
391 * fail if we can't post all our buffers at this time. */
395 /* Now allocate pool of reply buffers */
396 /* Increase max reply size to next power of two */
397 service->srv_max_reply_size = 1;
398 while (service->srv_max_reply_size <
399 max_reply_size + SPTLRPC_MAX_PAYLOAD)
400 service->srv_max_reply_size <<= 1;
402 if (proc_entry != NULL)
403 ptlrpc_lprocfs_register_service(proc_entry, service);
405 CDEBUG(D_NET, "%s: Started, listening on portal %d\n",
406 service->srv_name, service->srv_req_portal);
410 ptlrpc_unregister_service(service);
415 * to actually free the request, must be called without holding svc_lock.
416 * note it's caller's responsibility to unlink req->rq_list.
418 static void ptlrpc_server_free_request(struct ptlrpc_request *req)
420 LASSERT(atomic_read(&req->rq_refcount) == 0);
421 LASSERT(list_empty(&req->rq_timed_list));
423 /* DEBUG_REQ() assumes the reply state of a request with a valid
424 * ref will not be destroyed until that reference is dropped. */
425 ptlrpc_req_drop_rs(req);
427 sptlrpc_svc_ctx_decref(req);
429 if (req != &req->rq_rqbd->rqbd_req) {
430 /* NB request buffers use an embedded
431 * req if the incoming req unlinked the
432 * MD; this isn't one of them! */
433 OBD_FREE(req, sizeof(*req));
438 * drop a reference count of the request. if it reaches 0, we either
439 * put it into history list, or free it immediately.
441 static void ptlrpc_server_drop_request(struct ptlrpc_request *req)
443 struct ptlrpc_request_buffer_desc *rqbd = req->rq_rqbd;
444 struct ptlrpc_service *svc = rqbd->rqbd_service;
446 struct list_head *tmp;
447 struct list_head *nxt;
449 if (!atomic_dec_and_test(&req->rq_refcount))
452 spin_lock(&svc->srv_lock);
454 svc->srv_n_active_reqs--;
455 list_add(&req->rq_list, &rqbd->rqbd_reqs);
457 refcount = --(rqbd->rqbd_refcount);
459 /* request buffer is now idle: add to history */
460 list_del(&rqbd->rqbd_list);
461 list_add_tail(&rqbd->rqbd_list, &svc->srv_history_rqbds);
462 svc->srv_n_history_rqbds++;
464 /* cull some history?
465 * I expect only about 1 or 2 rqbds need to be recycled here */
466 while (svc->srv_n_history_rqbds > svc->srv_max_history_rqbds) {
467 rqbd = list_entry(svc->srv_history_rqbds.next,
468 struct ptlrpc_request_buffer_desc,
471 list_del(&rqbd->rqbd_list);
472 svc->srv_n_history_rqbds--;
474 /* remove rqbd's reqs from svc's req history while
475 * I've got the service lock */
476 list_for_each(tmp, &rqbd->rqbd_reqs) {
477 req = list_entry(tmp, struct ptlrpc_request,
479 /* Track the highest culled req seq */
480 if (req->rq_history_seq >
481 svc->srv_request_max_cull_seq)
482 svc->srv_request_max_cull_seq =
484 list_del(&req->rq_history_list);
487 spin_unlock(&svc->srv_lock);
489 list_for_each_safe(tmp, nxt, &rqbd->rqbd_reqs) {
490 req = list_entry(rqbd->rqbd_reqs.next,
491 struct ptlrpc_request,
493 list_del(&req->rq_list);
494 ptlrpc_server_free_request(req);
497 spin_lock(&svc->srv_lock);
499 * now all reqs including the embedded req has been
500 * disposed, schedule request buffer for re-use.
502 LASSERT(atomic_read(&rqbd->rqbd_req.rq_refcount) == 0);
503 list_add_tail(&rqbd->rqbd_list, &svc->srv_idle_rqbds);
506 spin_unlock(&svc->srv_lock);
507 } else if (req->rq_reply_state && req->rq_reply_state->rs_prealloc) {
508 /* If we are low on memory, we are not interested in history */
509 list_del(&req->rq_list);
510 list_del_init(&req->rq_history_list);
511 spin_unlock(&svc->srv_lock);
513 ptlrpc_server_free_request(req);
515 spin_unlock(&svc->srv_lock);
520 * to finish a request: stop sending more early replies, and release
521 * the request. should be called after we finished handling the request.
523 static void ptlrpc_server_finish_request(struct ptlrpc_request *req)
525 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
527 if (req->rq_export) {
528 class_export_put(req->rq_export);
529 req->rq_export = NULL;
532 if (req->rq_phase != RQ_PHASE_NEW) /* incorrect message magic */
533 DEBUG_REQ(D_INFO, req, "free req");
535 spin_lock(&svc->srv_at_lock);
536 req->rq_sent_final = 1;
537 list_del_init(&req->rq_timed_list);
538 spin_unlock(&svc->srv_at_lock);
540 ptlrpc_server_drop_request(req);
543 /* This function makes sure dead exports are evicted in a timely manner.
544 This function is only called when some export receives a message (i.e.,
545 the network is up.) */
546 static void ptlrpc_update_export_timer(struct obd_export *exp, long extra_delay)
548 struct obd_export *oldest_exp;
549 time_t oldest_time, new_time;
555 /* Compensate for slow machines, etc, by faking our request time
556 into the future. Although this can break the strict time-ordering
557 of the list, we can be really lazy here - we don't have to evict
558 at the exact right moment. Eventually, all silent exports
559 will make it to the top of the list. */
561 /* Do not pay attention on 1sec or smaller renewals. */
562 new_time = cfs_time_current_sec() + extra_delay;
563 if (exp->exp_last_request_time + 1 /*second */ >= new_time)
566 exp->exp_last_request_time = new_time;
567 CDEBUG(D_HA, "updating export %s at "CFS_TIME_T" exp %p\n",
568 exp->exp_client_uuid.uuid,
569 exp->exp_last_request_time, exp);
571 /* exports may get disconnected from the chain even though the
572 export has references, so we must keep the spin lock while
573 manipulating the lists */
574 spin_lock(&exp->exp_obd->obd_dev_lock);
576 if (list_empty(&exp->exp_obd_chain_timed)) {
577 /* this one is not timed */
578 spin_unlock(&exp->exp_obd->obd_dev_lock);
582 list_move_tail(&exp->exp_obd_chain_timed,
583 &exp->exp_obd->obd_exports_timed);
585 oldest_exp = list_entry(exp->exp_obd->obd_exports_timed.next,
586 struct obd_export, exp_obd_chain_timed);
587 oldest_time = oldest_exp->exp_last_request_time;
588 spin_unlock(&exp->exp_obd->obd_dev_lock);
590 if (exp->exp_obd->obd_recovering) {
591 /* be nice to everyone during recovery */
596 /* Note - racing to start/reset the obd_eviction timer is safe */
597 if (exp->exp_obd->obd_eviction_timer == 0) {
598 /* Check if the oldest entry is expired. */
599 if (cfs_time_current_sec() > (oldest_time + PING_EVICT_TIMEOUT +
601 /* We need a second timer, in case the net was down and
602 * it just came back. Since the pinger may skip every
603 * other PING_INTERVAL (see note in ptlrpc_pinger_main),
604 * we better wait for 3. */
605 exp->exp_obd->obd_eviction_timer =
606 cfs_time_current_sec() + 3 * PING_INTERVAL;
607 CDEBUG(D_HA, "%s: Think about evicting %s from "CFS_TIME_T"\n",
608 exp->exp_obd->obd_name, obd_export_nid2str(exp),
612 if (cfs_time_current_sec() >
613 (exp->exp_obd->obd_eviction_timer + extra_delay)) {
614 /* The evictor won't evict anyone who we've heard from
615 * recently, so we don't have to check before we start
617 if (!ping_evictor_wake(exp))
618 exp->exp_obd->obd_eviction_timer = 0;
625 static int ptlrpc_check_req(struct ptlrpc_request *req)
627 if (unlikely(lustre_msg_get_conn_cnt(req->rq_reqmsg) <
628 req->rq_export->exp_conn_cnt)) {
629 DEBUG_REQ(D_ERROR, req,
630 "DROPPING req from old connection %d < %d",
631 lustre_msg_get_conn_cnt(req->rq_reqmsg),
632 req->rq_export->exp_conn_cnt);
635 if (unlikely(req->rq_export->exp_obd &&
636 req->rq_export->exp_obd->obd_fail)) {
637 /* Failing over, don't handle any more reqs, send
638 error response instead. */
639 CDEBUG(D_RPCTRACE, "Dropping req %p for failed obd %s\n",
640 req, req->rq_export->exp_obd->obd_name);
641 req->rq_status = -ENODEV;
649 static void ptlrpc_at_set_timer(struct ptlrpc_service *svc)
651 struct ptlrpc_request *rq;
654 spin_lock(&svc->srv_at_lock);
655 if (list_empty(&svc->srv_at_list)) {
656 cfs_timer_disarm(&svc->srv_at_timer);
657 spin_unlock(&svc->srv_at_lock);
661 /* Set timer for closest deadline */
662 rq = list_entry(svc->srv_at_list.next, struct ptlrpc_request,
664 next = (__s32)(rq->rq_deadline - cfs_time_current_sec() -
667 ptlrpc_at_timer((unsigned long)svc);
669 cfs_timer_arm(&svc->srv_at_timer, cfs_time_shift(next));
670 spin_unlock(&svc->srv_at_lock);
671 CDEBUG(D_INFO, "armed %s at %+ds\n", svc->srv_name, next);
674 /* Add rpc to early reply check list */
675 static int ptlrpc_at_add_timed(struct ptlrpc_request *req)
677 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
678 struct ptlrpc_request *rq;
684 if (req->rq_no_reply)
687 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0)
690 spin_lock(&svc->srv_at_lock);
692 if (unlikely(req->rq_sent_final)) {
693 spin_unlock(&svc->srv_at_lock);
697 LASSERT(list_empty(&req->rq_timed_list));
698 /* Add to sorted list. Presumably latest rpcs will have the latest
699 deadlines, so search backward. */
700 list_for_each_entry_reverse(rq, &svc->srv_at_list, rq_timed_list) {
701 if (req->rq_deadline >= rq->rq_deadline) {
702 list_add(&req->rq_timed_list, &rq->rq_timed_list);
708 /* Add to front if shortest deadline or list empty */
709 list_add(&req->rq_timed_list, &svc->srv_at_list);
711 /* Check if we're the head of the list */
712 found = (svc->srv_at_list.next == &req->rq_timed_list);
714 spin_unlock(&svc->srv_at_lock);
717 ptlrpc_at_set_timer(svc);
722 static int ptlrpc_at_send_early_reply(struct ptlrpc_request *req,
725 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
726 struct ptlrpc_request *reqcopy;
727 struct lustre_msg *reqmsg;
728 cfs_duration_t olddl = req->rq_deadline - cfs_time_current_sec();
733 /* deadline is when the client expects us to reply, margin is the
734 difference between clients' and servers' expectations */
735 DEBUG_REQ(D_ADAPTTO, req,
736 "%ssending early reply (deadline %+lds, margin %+lds) for "
737 "%d+%d", AT_OFF ? "AT off - not " : "",
738 olddl, olddl - at_get(&svc->srv_at_estimate),
739 at_get(&svc->srv_at_estimate), extra_time);
745 DEBUG_REQ(D_WARNING, req, "Already past deadline (%+lds), "
746 "not sending early reply. Consider increasing "
747 "at_early_margin (%d)?", olddl, at_early_margin);
749 /* Return an error so we're not re-added to the timed list. */
753 if ((lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT) == 0){
754 DEBUG_REQ(D_INFO, req, "Wanted to ask client for more time, "
755 "but no AT support");
759 if (req->rq_export && req->rq_export->exp_in_recovery) {
760 /* don't increase server estimates during recovery, and give
761 clients the full recovery time. */
762 newdl = cfs_time_current_sec() +
763 req->rq_export->exp_obd->obd_recovery_timeout;
766 /* Fake our processing time into the future to ask the
767 clients for some extra amount of time */
768 extra_time += cfs_time_current_sec() -
769 req->rq_arrival_time.tv_sec;
770 at_add(&svc->srv_at_estimate, extra_time);
772 newdl = req->rq_arrival_time.tv_sec +
773 at_get(&svc->srv_at_estimate);
775 if (req->rq_deadline >= newdl) {
776 /* We're not adding any time, no need to send an early reply
777 (e.g. maybe at adaptive_max) */
778 DEBUG_REQ(D_WARNING, req, "Couldn't add any time ("
779 CFS_DURATION_T"/"CFS_DURATION_T"), "
780 "not sending early reply\n", olddl,
781 cfs_time_sub(newdl, cfs_time_current_sec()));
785 OBD_ALLOC(reqcopy, sizeof *reqcopy);
788 OBD_ALLOC(reqmsg, req->rq_reqlen);
790 OBD_FREE(reqcopy, sizeof *reqcopy);
795 reqcopy->rq_reply_state = NULL;
796 reqcopy->rq_rep_swab_mask = 0;
797 reqcopy->rq_pack_bulk = 0;
798 reqcopy->rq_pack_udesc = 0;
799 reqcopy->rq_packed_final = 0;
800 sptlrpc_svc_ctx_addref(reqcopy);
801 /* We only need the reqmsg for the magic */
802 reqcopy->rq_reqmsg = reqmsg;
803 memcpy(reqmsg, req->rq_reqmsg, req->rq_reqlen);
805 if (req->rq_sent_final) {
806 DEBUG_REQ(D_ADAPTTO, reqcopy, "Normal reply already sent out, "
807 "abort sending early reply\n");
812 reqcopy->rq_export = class_conn2export(
813 lustre_msg_get_handle(reqcopy->rq_reqmsg));
814 if (reqcopy->rq_export == NULL)
815 GOTO(out, rc = -ENODEV);
818 class_export_rpc_get(reqcopy->rq_export);
819 if (reqcopy->rq_export->exp_obd &&
820 reqcopy->rq_export->exp_obd->obd_fail)
821 GOTO(out_put, rc = -ENODEV);
823 rc = lustre_pack_reply_flags(reqcopy, 1, NULL, NULL, LPRFL_EARLY_REPLY);
827 rc = ptlrpc_send_reply(reqcopy, PTLRPC_REPLY_EARLY);
830 /* Adjust our own deadline to what we told the client */
831 req->rq_deadline = newdl;
832 req->rq_early_count++; /* number sent, server side */
834 DEBUG_REQ(D_ERROR, req, "Early reply send failed %d", rc);
837 /* Free the (early) reply state from lustre_pack_reply.
838 (ptlrpc_send_reply takes it's own rs ref, so this is safe here) */
839 ptlrpc_req_drop_rs(reqcopy);
842 class_export_rpc_put(reqcopy->rq_export);
843 class_export_put(reqcopy->rq_export);
845 sptlrpc_svc_ctx_decref(reqcopy);
846 OBD_FREE(reqmsg, req->rq_reqlen);
847 OBD_FREE(reqcopy, sizeof *reqcopy);
851 /* Send early replies to everybody expiring within at_early_margin
852 asking for at_extra time */
853 static int ptlrpc_at_check_timed(struct ptlrpc_service *svc)
855 struct ptlrpc_request *rq, *n;
856 struct list_head work_list;
857 time_t now = cfs_time_current_sec();
858 cfs_duration_t delay;
859 int first, counter = 0;
862 spin_lock(&svc->srv_at_lock);
863 if (svc->srv_at_check == 0) {
864 spin_unlock(&svc->srv_at_lock);
867 delay = cfs_time_sub(cfs_time_current(), svc->srv_at_checktime);
868 svc->srv_at_check = 0;
870 if (list_empty(&svc->srv_at_list)) {
871 spin_unlock(&svc->srv_at_lock);
875 /* The timer went off, but maybe the nearest rpc already completed. */
876 rq = list_entry(svc->srv_at_list.next, struct ptlrpc_request,
878 first = (int)(rq->rq_deadline - now);
879 if (first > at_early_margin) {
880 /* We've still got plenty of time. Reset the timer. */
881 spin_unlock(&svc->srv_at_lock);
882 ptlrpc_at_set_timer(svc);
886 /* We're close to a timeout, and we don't know how much longer the
887 server will take. Send early replies to everyone expiring soon. */
888 CFS_INIT_LIST_HEAD(&work_list);
889 list_for_each_entry_safe(rq, n, &svc->srv_at_list, rq_timed_list) {
890 if (rq->rq_deadline <= now + at_early_margin) {
891 list_move_tail(&rq->rq_timed_list, &work_list);
898 spin_unlock(&svc->srv_at_lock);
900 /* we have a new earliest deadline, restart the timer */
901 ptlrpc_at_set_timer(svc);
903 CDEBUG(D_ADAPTTO, "timeout in %+ds, asking for %d secs on %d early "
904 "replies\n", first, at_extra, counter);
906 /* We're already past request deadlines before we even get a
907 chance to send early replies */
908 LCONSOLE_WARN("%s: This server is not able to keep up with "
909 "request traffic (cpu-bound).\n", svc->srv_name);
910 CWARN("earlyQ=%d reqQ=%d recA=%d, svcEst=%d, "
911 "delay="CFS_DURATION_T"(jiff)\n",
912 counter, svc->srv_n_queued_reqs, svc->srv_n_active_reqs,
913 at_get(&svc->srv_at_estimate), delay);
916 /* ptlrpc_server_finish_request may delete an entry out of
918 spin_lock(&svc->srv_at_lock);
919 while (!list_empty(&work_list)) {
920 rq = list_entry(work_list.next, struct ptlrpc_request,
922 list_del_init(&rq->rq_timed_list);
923 /* if the entry is still in the worklist, it hasn't been
924 deleted, and is safe to take a ref to keep the req around */
925 atomic_inc(&rq->rq_refcount);
926 spin_unlock(&svc->srv_at_lock);
928 if (ptlrpc_at_send_early_reply(rq, at_extra) == 0)
929 ptlrpc_at_add_timed(rq);
931 ptlrpc_server_drop_request(rq);
932 spin_lock(&svc->srv_at_lock);
934 spin_unlock(&svc->srv_at_lock);
940 * Put the request to the export list if the request may become
941 * a high priority one.
943 static int ptlrpc_hpreq_init(struct ptlrpc_service *svc,
944 struct ptlrpc_request *req)
949 if (svc->srv_hpreq_handler) {
950 rc = svc->srv_hpreq_handler(req);
954 if (req->rq_export && req->rq_ops) {
955 spin_lock(&req->rq_export->exp_lock);
956 list_add(&req->rq_exp_list, &req->rq_export->exp_queued_rpc);
957 spin_unlock(&req->rq_export->exp_lock);
963 /** Remove the request from the export list. */
964 static void ptlrpc_hpreq_fini(struct ptlrpc_request *req)
967 if (req->rq_export && req->rq_ops) {
968 spin_lock(&req->rq_export->exp_lock);
969 list_del_init(&req->rq_exp_list);
970 spin_unlock(&req->rq_export->exp_lock);
976 * Make the request a high priority one.
978 * All the high priority requests are queued in a separate FIFO
979 * ptlrpc_service::srv_request_hpq list which is parallel to
980 * ptlrpc_service::srv_request_queue list but has a higher priority
983 * \see ptlrpc_server_handle_request().
985 static void ptlrpc_hpreq_reorder_nolock(struct ptlrpc_service *svc,
986 struct ptlrpc_request *req)
989 LASSERT(svc != NULL);
990 spin_lock(&req->rq_lock);
991 if (req->rq_hp == 0) {
992 int opc = lustre_msg_get_opc(req->rq_reqmsg);
994 /* Add to the high priority queue. */
995 list_move_tail(&req->rq_list, &svc->srv_request_hpq);
998 DEBUG_REQ(D_NET, req, "high priority req");
1000 spin_unlock(&req->rq_lock);
1004 void ptlrpc_hpreq_reorder(struct ptlrpc_request *req)
1006 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
1009 spin_lock(&svc->srv_lock);
1010 /* It may happen that the request is already taken for the processing
1011 * but still in the export list, do not re-add it into the HP list. */
1012 if (req->rq_phase == RQ_PHASE_NEW)
1013 ptlrpc_hpreq_reorder_nolock(svc, req);
1014 spin_unlock(&svc->srv_lock);
1018 /** Check if the request if a high priority one. */
1019 static int ptlrpc_server_hpreq_check(struct ptlrpc_request *req)
1024 /* Check by request opc. */
1025 opc = lustre_msg_get_opc(req->rq_reqmsg);
1026 if (opc == OBD_PING)
1029 /* Perform request specific check. */
1030 if (req->rq_ops && req->rq_ops->hpreq_check)
1031 rc = req->rq_ops->hpreq_check(req);
1035 /** Check if a request is a high priority one. */
1036 static int ptlrpc_server_request_add(struct ptlrpc_service *svc,
1037 struct ptlrpc_request *req)
1042 rc = ptlrpc_server_hpreq_check(req);
1046 spin_lock(&svc->srv_lock);
1047 /* Before inserting the request into the queue, check if it is not
1048 * inserted yet, or even already handled -- it may happen due to
1049 * a racing ldlm_server_blocking_ast(). */
1050 if (req->rq_phase == RQ_PHASE_NEW && list_empty(&req->rq_list)) {
1052 ptlrpc_hpreq_reorder_nolock(svc, req);
1054 list_add_tail(&req->rq_list, &svc->srv_request_queue);
1056 spin_unlock(&svc->srv_lock);
1061 /* Only allow normal priority requests on a service that has a high-priority
1062 * queue if forced (i.e. cleanup), if there are other high priority requests
1063 * already being processed (i.e. those threads can service more high-priority
1064 * requests), or if there are enough idle threads that a later thread can do
1065 * a high priority request. */
1066 static int ptlrpc_server_allow_normal(struct ptlrpc_service *svc, int force)
1068 return force || !svc->srv_hpreq_handler || svc->srv_n_hpreq > 0 ||
1069 svc->srv_n_active_reqs < svc->srv_threads_running - 2;
1072 static struct ptlrpc_request *
1073 ptlrpc_server_request_get(struct ptlrpc_service *svc, int force)
1075 struct ptlrpc_request *req = NULL;
1078 if (ptlrpc_server_allow_normal(svc, force) &&
1079 !list_empty(&svc->srv_request_queue) &&
1080 (list_empty(&svc->srv_request_hpq) ||
1081 svc->srv_hpreq_count >= svc->srv_hpreq_ratio)) {
1082 req = list_entry(svc->srv_request_queue.next,
1083 struct ptlrpc_request, rq_list);
1084 svc->srv_hpreq_count = 0;
1085 } else if (!list_empty(&svc->srv_request_hpq)) {
1086 req = list_entry(svc->srv_request_hpq.next,
1087 struct ptlrpc_request, rq_list);
1088 svc->srv_hpreq_count++;
1093 static int ptlrpc_server_request_pending(struct ptlrpc_service *svc, int force)
1095 return ((ptlrpc_server_allow_normal(svc, force) &&
1096 !list_empty(&svc->srv_request_queue)) ||
1097 !list_empty(&svc->srv_request_hpq));
1100 /* Handle freshly incoming reqs, add to timed early reply list,
1101 pass on to regular request queue */
1103 ptlrpc_server_handle_req_in(struct ptlrpc_service *svc)
1105 struct ptlrpc_request *req;
1112 spin_lock(&svc->srv_lock);
1113 if (list_empty(&svc->srv_req_in_queue)) {
1114 spin_unlock(&svc->srv_lock);
1118 req = list_entry(svc->srv_req_in_queue.next,
1119 struct ptlrpc_request, rq_list);
1120 list_del_init (&req->rq_list);
1121 /* Consider this still a "queued" request as far as stats are
1123 spin_unlock(&svc->srv_lock);
1125 /* go through security check/transform */
1126 rc = sptlrpc_svc_unwrap_request(req);
1130 case SECSVC_COMPLETE:
1131 target_send_reply(req, 0, OBD_FAIL_MDS_ALL_REPLY_NET);
1139 /* Clear request swab mask; this is a new request */
1140 req->rq_req_swab_mask = 0;
1142 rc = lustre_unpack_msg(req->rq_reqmsg, req->rq_reqlen);
1144 CERROR("error unpacking request: ptl %d from %s x"LPU64"\n",
1145 svc->srv_req_portal, libcfs_id2str(req->rq_peer),
1150 rc = lustre_unpack_req_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
1152 CERROR ("error unpacking ptlrpc body: ptl %d from %s x"
1153 LPU64"\n", svc->srv_req_portal,
1154 libcfs_id2str(req->rq_peer), req->rq_xid);
1159 if (lustre_msg_get_type(req->rq_reqmsg) != PTL_RPC_MSG_REQUEST) {
1160 CERROR("wrong packet type received (type=%u) from %s\n",
1161 lustre_msg_get_type(req->rq_reqmsg),
1162 libcfs_id2str(req->rq_peer));
1166 CDEBUG(D_NET, "got req "LPD64"\n", req->rq_xid);
1168 req->rq_export = class_conn2export(
1169 lustre_msg_get_handle(req->rq_reqmsg));
1170 if (req->rq_export) {
1171 rc = ptlrpc_check_req(req);
1173 rc = sptlrpc_target_export_check(req->rq_export, req);
1175 DEBUG_REQ(D_ERROR, req, "DROPPING req with "
1176 "illegal security flavor,");
1181 ptlrpc_update_export_timer(req->rq_export, 0);
1184 /* req_in handling should/must be fast */
1185 if (cfs_time_current_sec() - req->rq_arrival_time.tv_sec > 5)
1186 DEBUG_REQ(D_WARNING, req, "Slow req_in handling "CFS_DURATION_T"s",
1187 cfs_time_sub(cfs_time_current_sec(),
1188 req->rq_arrival_time.tv_sec));
1190 /* Set rpc server deadline and add it to the timed list */
1191 deadline = (lustre_msghdr_get_flags(req->rq_reqmsg) &
1192 MSGHDR_AT_SUPPORT) ?
1193 /* The max time the client expects us to take */
1194 lustre_msg_get_timeout(req->rq_reqmsg) : obd_timeout;
1195 req->rq_deadline = req->rq_arrival_time.tv_sec + deadline;
1196 if (unlikely(deadline == 0)) {
1197 DEBUG_REQ(D_ERROR, req, "Dropping request with 0 timeout");
1201 ptlrpc_at_add_timed(req);
1202 rc = ptlrpc_hpreq_init(svc, req);
1206 /* Move it over to the request processing queue */
1207 rc = ptlrpc_server_request_add(svc, req);
1210 cfs_waitq_signal(&svc->srv_waitq);
1214 spin_lock(&svc->srv_lock);
1215 svc->srv_n_queued_reqs--;
1216 svc->srv_n_active_reqs++;
1217 spin_unlock(&svc->srv_lock);
1218 ptlrpc_server_finish_request(req);
1224 ptlrpc_server_handle_request(struct ptlrpc_service *svc,
1225 struct ptlrpc_thread *thread)
1227 struct obd_export *export = NULL;
1228 struct ptlrpc_request *request;
1229 struct timeval work_start;
1230 struct timeval work_end;
1238 spin_lock(&svc->srv_lock);
1239 if (unlikely(!ptlrpc_server_request_pending(svc, 0) ||
1242 /* !@%$# liblustre only has 1 thread */
1243 svc->srv_n_difficult_replies != 0 &&
1245 svc->srv_n_active_reqs >= (svc->srv_threads_running - 1)))) {
1246 /* Don't handle regular requests in the last thread, in order * re
1247 * to handle difficult replies (which might block other threads)
1248 * as well as handle any incoming reqs, early replies, etc.
1249 * That means we always need at least 2 service threads. */
1250 spin_unlock(&svc->srv_lock);
1254 request = ptlrpc_server_request_get(svc, 0);
1255 if (request == NULL) {
1256 spin_unlock(&svc->srv_lock);
1260 opc = lustre_msg_get_opc(request->rq_reqmsg);
1261 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT))
1262 fail_opc = OBD_FAIL_PTLRPC_HPREQ_NOTIMEOUT;
1263 else if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_HPREQ_TIMEOUT))
1264 fail_opc = OBD_FAIL_PTLRPC_HPREQ_TIMEOUT;
1266 if (unlikely(fail_opc)) {
1267 if (request->rq_export && request->rq_ops) {
1268 spin_unlock(&svc->srv_lock);
1269 OBD_FAIL_TIMEOUT(fail_opc, 4);
1270 spin_lock(&svc->srv_lock);
1271 request = ptlrpc_server_request_get(svc, 0);
1272 if (request == NULL) {
1273 spin_unlock(&svc->srv_lock);
1276 LASSERT(ptlrpc_server_request_pending(svc, 0));
1280 list_del_init(&request->rq_list);
1281 svc->srv_n_queued_reqs--;
1282 svc->srv_n_active_reqs++;
1286 /* The phase is changed under the lock here because we need to know
1287 * the request is under processing (see ptlrpc_hpreq_reorder()). */
1288 ptlrpc_rqphase_move(request, RQ_PHASE_INTERPRET);
1289 spin_unlock(&svc->srv_lock);
1291 ptlrpc_hpreq_fini(request);
1293 if(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DUMP_LOG))
1294 libcfs_debug_dumplog();
1296 do_gettimeofday(&work_start);
1297 timediff = cfs_timeval_sub(&work_start, &request->rq_arrival_time,NULL);
1298 if (likely(svc->srv_stats != NULL)) {
1299 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQWAIT_CNTR,
1301 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQQDEPTH_CNTR,
1302 svc->srv_n_queued_reqs);
1303 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQACTIVE_CNTR,
1304 svc->srv_n_active_reqs);
1305 lprocfs_counter_add(svc->srv_stats, PTLRPC_TIMEOUT,
1306 at_get(&svc->srv_at_estimate));
1309 rc = lu_context_init(&request->rq_session,
1310 LCT_SESSION|LCT_REMEMBER|LCT_NOREF);
1312 CERROR("Failure to initialize session: %d\n", rc);
1315 request->rq_session.lc_thread = thread;
1316 request->rq_session.lc_cookie = 0x5;
1317 lu_context_enter(&request->rq_session);
1319 CDEBUG(D_NET, "got req "LPU64"\n", request->rq_xid);
1321 request->rq_svc_thread = thread;
1323 request->rq_svc_thread->t_env->le_ses = &request->rq_session;
1325 if (likely(request->rq_export)) {
1326 if (unlikely(ptlrpc_check_req(request)))
1328 ptlrpc_update_export_timer(request->rq_export, timediff >> 19);
1329 export = class_export_rpc_get(request->rq_export);
1332 /* Discard requests queued for longer than the deadline.
1333 The deadline is increased if we send an early reply. */
1334 if (cfs_time_current_sec() > request->rq_deadline) {
1335 DEBUG_REQ(D_ERROR, request, "Dropping timed-out request from %s"
1336 ": deadline "CFS_DURATION_T":"CFS_DURATION_T"s ago\n",
1337 libcfs_id2str(request->rq_peer),
1338 cfs_time_sub(request->rq_deadline,
1339 request->rq_arrival_time.tv_sec),
1340 cfs_time_sub(cfs_time_current_sec(),
1341 request->rq_deadline));
1342 goto put_rpc_export;
1345 CDEBUG(D_RPCTRACE, "Handling RPC pname:cluuid+ref:pid:xid:nid:opc "
1346 "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1347 (request->rq_export ?
1348 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1349 (request->rq_export ?
1350 atomic_read(&request->rq_export->exp_refcount) : -99),
1351 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1352 libcfs_id2str(request->rq_peer),
1353 lustre_msg_get_opc(request->rq_reqmsg));
1355 OBD_FAIL_TIMEOUT_MS(OBD_FAIL_PTLRPC_PAUSE_REQ, obd_fail_val);
1357 rc = svc->srv_handler(request);
1359 ptlrpc_rqphase_move(request, RQ_PHASE_COMPLETE);
1361 CDEBUG(D_RPCTRACE, "Handled RPC pname:cluuid+ref:pid:xid:nid:opc "
1362 "%s:%s+%d:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1363 (request->rq_export ?
1364 (char *)request->rq_export->exp_client_uuid.uuid : "0"),
1365 (request->rq_export ?
1366 atomic_read(&request->rq_export->exp_refcount) : -99),
1367 lustre_msg_get_status(request->rq_reqmsg), request->rq_xid,
1368 libcfs_id2str(request->rq_peer),
1369 lustre_msg_get_opc(request->rq_reqmsg));
1373 class_export_rpc_put(export);
1375 lu_context_exit(&request->rq_session);
1376 lu_context_fini(&request->rq_session);
1378 if (unlikely(cfs_time_current_sec() > request->rq_deadline)) {
1379 DEBUG_REQ(D_WARNING, request, "Request x"LPU64" took longer "
1380 "than estimated ("CFS_DURATION_T":"CFS_DURATION_T"s);"
1381 " client may timeout.",
1382 request->rq_xid, cfs_time_sub(request->rq_deadline,
1383 request->rq_arrival_time.tv_sec),
1384 cfs_time_sub(cfs_time_current_sec(),
1385 request->rq_deadline));
1388 do_gettimeofday(&work_end);
1389 timediff = cfs_timeval_sub(&work_end, &work_start, NULL);
1390 CDEBUG(D_RPCTRACE, "request x"LPU64" opc %u from %s processed in "
1391 "%ldus (%ldus total) trans "LPU64" rc %d/%d\n",
1392 request->rq_xid, lustre_msg_get_opc(request->rq_reqmsg),
1393 libcfs_id2str(request->rq_peer), timediff,
1394 cfs_timeval_sub(&work_end, &request->rq_arrival_time, NULL),
1395 request->rq_repmsg ? lustre_msg_get_transno(request->rq_repmsg) :
1396 request->rq_transno, request->rq_status,
1397 request->rq_repmsg ? lustre_msg_get_status(request->rq_repmsg):
1399 if (likely(svc->srv_stats != NULL && request->rq_reqmsg != NULL)) {
1400 __u32 op = lustre_msg_get_opc(request->rq_reqmsg);
1401 int opc = opcode_offset(op);
1402 if (opc > 0 && !(op == LDLM_ENQUEUE || op == MDS_REINT)) {
1403 LASSERT(opc < LUSTRE_MAX_OPCODES);
1404 lprocfs_counter_add(svc->srv_stats,
1405 opc + EXTRA_MAX_OPCODES,
1409 if (unlikely(request->rq_early_count)) {
1410 DEBUG_REQ(D_ADAPTTO, request,
1411 "sent %d early replies before finishing in "
1413 request->rq_early_count,
1414 cfs_time_sub(work_end.tv_sec,
1415 request->rq_arrival_time.tv_sec));
1419 spin_lock(&svc->srv_lock);
1422 spin_unlock(&svc->srv_lock);
1423 ptlrpc_server_finish_request(request);
1429 ptlrpc_server_handle_reply (struct ptlrpc_service *svc)
1431 struct ptlrpc_reply_state *rs;
1432 struct obd_export *exp;
1433 struct obd_device *obd;
1438 spin_lock(&svc->srv_lock);
1439 if (list_empty (&svc->srv_reply_queue)) {
1440 spin_unlock(&svc->srv_lock);
1444 rs = list_entry (svc->srv_reply_queue.next,
1445 struct ptlrpc_reply_state, rs_list);
1447 exp = rs->rs_export;
1450 LASSERT (rs->rs_difficult);
1451 LASSERT (rs->rs_scheduled);
1453 list_del_init (&rs->rs_list);
1455 /* Disengage from notifiers carefully (lock order - irqrestore below!)*/
1456 spin_unlock(&svc->srv_lock);
1458 spin_lock (&obd->obd_uncommitted_replies_lock);
1459 /* Noop if removed already */
1460 list_del_init (&rs->rs_obd_list);
1461 spin_unlock (&obd->obd_uncommitted_replies_lock);
1463 spin_lock (&exp->exp_lock);
1464 /* Noop if removed already */
1465 list_del_init (&rs->rs_exp_list);
1466 spin_unlock (&exp->exp_lock);
1468 spin_lock(&svc->srv_lock);
1470 been_handled = rs->rs_handled;
1473 nlocks = rs->rs_nlocks; /* atomic "steal", but */
1474 rs->rs_nlocks = 0; /* locks still on rs_locks! */
1476 if (nlocks == 0 && !been_handled) {
1477 /* If we see this, we should already have seen the warning
1478 * in mds_steal_ack_locks() */
1479 CWARN("All locks stolen from rs %p x"LPD64".t"LPD64
1482 rs->rs_xid, rs->rs_transno,
1483 lustre_msg_get_opc(rs->rs_msg),
1484 libcfs_nid2str(exp->exp_connection->c_peer.nid));
1487 if ((!been_handled && rs->rs_on_net) || nlocks > 0) {
1488 spin_unlock(&svc->srv_lock);
1490 if (!been_handled && rs->rs_on_net) {
1491 LNetMDUnlink(rs->rs_md_h);
1492 /* Ignore return code; we're racing with
1496 while (nlocks-- > 0)
1497 ldlm_lock_decref(&rs->rs_locks[nlocks],
1498 rs->rs_modes[nlocks]);
1500 spin_lock(&svc->srv_lock);
1503 rs->rs_scheduled = 0;
1505 if (!rs->rs_on_net) {
1507 svc->srv_n_difficult_replies--;
1508 if (svc->srv_n_difficult_replies == 0 && svc->srv_is_stopping)
1509 /* wake up threads that are being stopped by
1510 ptlrpc_unregister_service/ptlrpc_stop_threads
1511 and sleep waiting svr_n_difficult_replies == 0 */
1512 cfs_waitq_broadcast(&svc->srv_waitq);
1513 spin_unlock(&svc->srv_lock);
1515 class_export_put (exp);
1516 rs->rs_export = NULL;
1517 ptlrpc_rs_decref (rs);
1518 atomic_dec (&svc->srv_outstanding_replies);
1522 /* still on the net; callback will schedule */
1523 spin_unlock(&svc->srv_lock);
1528 /* FIXME make use of timeout later */
1530 liblustre_check_services (void *arg)
1532 int did_something = 0;
1534 struct list_head *tmp, *nxt;
1537 /* I'm relying on being single threaded, not to have to lock
1538 * ptlrpc_all_services etc */
1539 list_for_each_safe (tmp, nxt, &ptlrpc_all_services) {
1540 struct ptlrpc_service *svc =
1541 list_entry (tmp, struct ptlrpc_service, srv_list);
1543 if (svc->srv_threads_running != 0) /* I've recursed */
1546 /* service threads can block for bulk, so this limits us
1547 * (arbitrarily) to recursing 1 stack frame per service.
1548 * Note that the problem with recursion is that we have to
1549 * unwind completely before our caller can resume. */
1551 svc->srv_threads_running++;
1554 rc = ptlrpc_server_handle_req_in(svc);
1555 rc |= ptlrpc_server_handle_reply(svc);
1556 rc |= ptlrpc_at_check_timed(svc);
1557 rc |= ptlrpc_server_handle_request(svc, NULL);
1558 rc |= (ptlrpc_server_post_idle_rqbds(svc) > 0);
1559 did_something |= rc;
1562 svc->srv_threads_running--;
1565 RETURN(did_something);
1567 #define ptlrpc_stop_all_threads(s) do {} while (0)
1569 #else /* __KERNEL__ */
1571 /* Don't use daemonize, it removes fs struct from new thread (bug 418) */
1572 void ptlrpc_daemonize(char *name)
1574 struct fs_struct *fs = current->fs;
1576 atomic_inc(&fs->count);
1577 cfs_daemonize(name);
1578 exit_fs(cfs_current());
1580 ll_set_fs_pwd(current->fs, init_task.fs->pwdmnt, init_task.fs->pwd);
1584 ptlrpc_check_rqbd_pool(struct ptlrpc_service *svc)
1586 int avail = svc->srv_nrqbd_receiving;
1587 int low_water = test_req_buffer_pressure ? 0 :
1588 svc->srv_nbuf_per_group/2;
1590 /* NB I'm not locking; just looking. */
1592 /* CAVEAT EMPTOR: We might be allocating buffers here because we've
1593 * allowed the request history to grow out of control. We could put a
1594 * sanity check on that here and cull some history if we need the
1597 if (avail <= low_water)
1598 ptlrpc_grow_req_bufs(svc);
1601 lprocfs_counter_add(svc->srv_stats, PTLRPC_REQBUF_AVAIL_CNTR,
1606 ptlrpc_retry_rqbds(void *arg)
1608 struct ptlrpc_service *svc = (struct ptlrpc_service *)arg;
1610 svc->srv_rqbd_timeout = 0;
1611 return (-ETIMEDOUT);
1614 static int ptlrpc_main(void *arg)
1616 struct ptlrpc_svc_data *data = (struct ptlrpc_svc_data *)arg;
1617 struct ptlrpc_service *svc = data->svc;
1618 struct ptlrpc_thread *thread = data->thread;
1619 struct obd_device *dev = data->dev;
1620 struct ptlrpc_reply_state *rs;
1621 #ifdef WITH_GROUP_INFO
1622 struct group_info *ginfo = NULL;
1625 int counter = 0, rc = 0;
1628 ptlrpc_daemonize(data->name);
1630 #if defined(HAVE_NODE_TO_CPUMASK) && defined(CONFIG_NUMA)
1631 /* we need to do this before any per-thread allocation is done so that
1632 * we get the per-thread allocations on local node. bug 7342 */
1633 if (svc->srv_cpu_affinity) {
1636 for (cpu = 0, num_cpu = 0; cpu < num_possible_cpus(); cpu++) {
1637 if (!cpu_online(cpu))
1639 if (num_cpu == thread->t_id % num_online_cpus())
1643 set_cpus_allowed(cfs_current(), node_to_cpumask(cpu_to_node(cpu)));
1647 #ifdef WITH_GROUP_INFO
1648 ginfo = groups_alloc(0);
1654 set_current_groups(ginfo);
1655 put_group_info(ginfo);
1658 if (svc->srv_init != NULL) {
1659 rc = svc->srv_init(thread);
1664 rc = lu_context_init(&env.le_ctx,
1665 svc->srv_ctx_tags|LCT_REMEMBER|LCT_NOREF);
1669 thread->t_env = &env;
1670 env.le_ctx.lc_thread = thread;
1671 env.le_ctx.lc_cookie = 0x6;
1673 /* Alloc reply state structure for this one */
1674 OBD_ALLOC_GFP(rs, svc->srv_max_reply_size, CFS_ALLOC_STD);
1680 /* Record that the thread is running */
1681 thread->t_flags = SVC_RUNNING;
1683 * wake up our creator. Note: @data is invalid after this point,
1684 * because it's allocated on ptlrpc_start_thread() stack.
1686 cfs_waitq_signal(&thread->t_ctl_waitq);
1688 thread->t_watchdog = lc_watchdog_add(max_t(int, obd_timeout, AT_OFF ? 0 :
1689 at_get(&svc->srv_at_estimate))
1690 * svc->srv_watchdog_factor,
1693 spin_lock(&svc->srv_lock);
1694 svc->srv_threads_running++;
1695 list_add(&rs->rs_list, &svc->srv_free_rs_list);
1696 spin_unlock(&svc->srv_lock);
1697 cfs_waitq_signal(&svc->srv_free_rs_waitq);
1699 CDEBUG(D_NET, "service thread %d (#%d) started\n", thread->t_id,
1700 svc->srv_threads_running);
1702 /* XXX maintain a list of all managed devices: insert here */
1704 while ((thread->t_flags & SVC_STOPPING) == 0 ||
1705 svc->srv_n_difficult_replies != 0) {
1706 /* Don't exit while there are replies to be handled */
1707 struct l_wait_info lwi = LWI_TIMEOUT(svc->srv_rqbd_timeout,
1708 ptlrpc_retry_rqbds, svc);
1710 lc_watchdog_disable(thread->t_watchdog);
1714 l_wait_event_exclusive (svc->srv_waitq,
1715 ((thread->t_flags & SVC_STOPPING) != 0 &&
1716 svc->srv_n_difficult_replies == 0) ||
1717 (!list_empty(&svc->srv_idle_rqbds) &&
1718 svc->srv_rqbd_timeout == 0) ||
1719 !list_empty(&svc->srv_req_in_queue) ||
1720 !list_empty(&svc->srv_reply_queue) ||
1721 (ptlrpc_server_request_pending(svc, 0) &&
1722 (svc->srv_n_active_reqs <
1723 (svc->srv_threads_running - 1))) ||
1727 lc_watchdog_touch_ms(thread->t_watchdog, max_t(int, obd_timeout,
1729 at_get(&svc->srv_at_estimate)) *
1730 svc->srv_watchdog_factor);
1732 ptlrpc_check_rqbd_pool(svc);
1734 if ((svc->srv_threads_started < svc->srv_threads_max) &&
1735 (svc->srv_n_active_reqs >= (svc->srv_threads_started - 1))){
1736 /* Ignore return code - we tried... */
1737 ptlrpc_start_thread(dev, svc);
1740 if (!list_empty(&svc->srv_reply_queue))
1741 ptlrpc_server_handle_reply(svc);
1743 if (!list_empty(&svc->srv_req_in_queue)) {
1744 /* Process all incoming reqs before handling any */
1745 ptlrpc_server_handle_req_in(svc);
1746 /* but limit ourselves in case of flood */
1747 if (counter++ < 1000)
1752 if (svc->srv_at_check)
1753 ptlrpc_at_check_timed(svc);
1755 /* don't handle requests in the last thread */
1756 if (ptlrpc_server_request_pending(svc, 0) &&
1757 (svc->srv_n_active_reqs < (svc->srv_threads_running - 1))) {
1758 lu_context_enter(&env.le_ctx);
1759 ptlrpc_server_handle_request(svc, thread);
1760 lu_context_exit(&env.le_ctx);
1763 if (!list_empty(&svc->srv_idle_rqbds) &&
1764 ptlrpc_server_post_idle_rqbds(svc) < 0) {
1765 /* I just failed to repost request buffers. Wait
1766 * for a timeout (unless something else happens)
1767 * before I try again */
1768 svc->srv_rqbd_timeout = cfs_time_seconds(1)/10;
1769 CDEBUG(D_RPCTRACE,"Posted buffers: %d\n",
1770 svc->srv_nrqbd_receiving);
1774 lc_watchdog_delete(thread->t_watchdog);
1775 thread->t_watchdog = NULL;
1779 * deconstruct service specific state created by ptlrpc_start_thread()
1781 if (svc->srv_done != NULL)
1782 svc->srv_done(thread);
1784 lu_context_fini(&env.le_ctx);
1786 CDEBUG(D_NET, "service thread %d exiting: rc %d\n", thread->t_id, rc);
1788 spin_lock(&svc->srv_lock);
1789 svc->srv_threads_running--; /* must know immediately */
1791 thread->t_flags = SVC_STOPPED;
1793 cfs_waitq_signal(&thread->t_ctl_waitq);
1794 spin_unlock(&svc->srv_lock);
1799 static void ptlrpc_stop_thread(struct ptlrpc_service *svc,
1800 struct ptlrpc_thread *thread)
1802 struct l_wait_info lwi = { 0 };
1805 CDEBUG(D_RPCTRACE, "Stopping thread %p\n", thread);
1806 spin_lock(&svc->srv_lock);
1807 thread->t_flags = SVC_STOPPING;
1808 spin_unlock(&svc->srv_lock);
1810 cfs_waitq_broadcast(&svc->srv_waitq);
1811 l_wait_event(thread->t_ctl_waitq, (thread->t_flags & SVC_STOPPED),
1814 spin_lock(&svc->srv_lock);
1815 list_del(&thread->t_link);
1816 spin_unlock(&svc->srv_lock);
1818 OBD_FREE_PTR(thread);
1822 void ptlrpc_stop_all_threads(struct ptlrpc_service *svc)
1824 struct ptlrpc_thread *thread;
1827 spin_lock(&svc->srv_lock);
1828 while (!list_empty(&svc->srv_threads)) {
1829 thread = list_entry(svc->srv_threads.next,
1830 struct ptlrpc_thread, t_link);
1832 spin_unlock(&svc->srv_lock);
1833 ptlrpc_stop_thread(svc, thread);
1834 spin_lock(&svc->srv_lock);
1837 spin_unlock(&svc->srv_lock);
1841 int ptlrpc_start_threads(struct obd_device *dev, struct ptlrpc_service *svc)
1846 /* We require 2 threads min - see note in
1847 ptlrpc_server_handle_request */
1848 LASSERT(svc->srv_threads_min >= 2);
1849 for (i = 0; i < svc->srv_threads_min; i++) {
1850 rc = ptlrpc_start_thread(dev, svc);
1851 /* We have enough threads, don't start more. b=15759 */
1855 CERROR("cannot start %s thread #%d: rc %d\n",
1856 svc->srv_thread_name, i, rc);
1857 ptlrpc_stop_all_threads(svc);
1863 int ptlrpc_start_thread(struct obd_device *dev, struct ptlrpc_service *svc)
1865 struct l_wait_info lwi = { 0 };
1866 struct ptlrpc_svc_data d;
1867 struct ptlrpc_thread *thread;
1872 CDEBUG(D_RPCTRACE, "%s started %d min %d max %d running %d\n",
1873 svc->srv_name, svc->srv_threads_started, svc->srv_threads_min,
1874 svc->srv_threads_max, svc->srv_threads_running);
1875 if (unlikely(svc->srv_threads_started >= svc->srv_threads_max) ||
1876 (OBD_FAIL_CHECK(OBD_FAIL_TGT_TOOMANY_THREADS) &&
1877 svc->srv_threads_started == svc->srv_threads_min - 1))
1880 OBD_ALLOC_PTR(thread);
1883 cfs_waitq_init(&thread->t_ctl_waitq);
1885 spin_lock(&svc->srv_lock);
1886 if (svc->srv_threads_started >= svc->srv_threads_max) {
1887 spin_unlock(&svc->srv_lock);
1888 OBD_FREE_PTR(thread);
1891 list_add(&thread->t_link, &svc->srv_threads);
1892 id = svc->srv_threads_started++;
1893 spin_unlock(&svc->srv_lock);
1896 sprintf(name, "%s_%02d", svc->srv_thread_name, id);
1902 CDEBUG(D_RPCTRACE, "starting thread '%s'\n", name);
1904 /* CLONE_VM and CLONE_FILES just avoid a needless copy, because we
1905 * just drop the VM and FILES in ptlrpc_daemonize() right away.
1907 rc = cfs_kernel_thread(ptlrpc_main, &d, CLONE_VM | CLONE_FILES);
1909 CERROR("cannot start thread '%s': rc %d\n", name, rc);
1911 spin_lock(&svc->srv_lock);
1912 list_del(&thread->t_link);
1913 --svc->srv_threads_started;
1914 spin_unlock(&svc->srv_lock);
1916 OBD_FREE(thread, sizeof(*thread));
1919 l_wait_event(thread->t_ctl_waitq,
1920 thread->t_flags & (SVC_RUNNING | SVC_STOPPED), &lwi);
1922 rc = (thread->t_flags & SVC_STOPPED) ? thread->t_id : 0;
1927 int ptlrpc_unregister_service(struct ptlrpc_service *service)
1930 struct l_wait_info lwi;
1931 struct list_head *tmp;
1932 struct ptlrpc_reply_state *rs, *t;
1935 service->srv_is_stopping = 1;
1936 cfs_timer_disarm(&service->srv_at_timer);
1938 ptlrpc_stop_all_threads(service);
1939 LASSERT(list_empty(&service->srv_threads));
1941 spin_lock (&ptlrpc_all_services_lock);
1942 list_del_init (&service->srv_list);
1943 spin_unlock (&ptlrpc_all_services_lock);
1945 ptlrpc_lprocfs_unregister_service(service);
1947 /* All history will be culled when the next request buffer is
1949 service->srv_max_history_rqbds = 0;
1951 CDEBUG(D_NET, "%s: tearing down\n", service->srv_name);
1953 rc = LNetClearLazyPortal(service->srv_req_portal);
1956 /* Unlink all the request buffers. This forces a 'final' event with
1957 * its 'unlink' flag set for each posted rqbd */
1958 list_for_each(tmp, &service->srv_active_rqbds) {
1959 struct ptlrpc_request_buffer_desc *rqbd =
1960 list_entry(tmp, struct ptlrpc_request_buffer_desc,
1963 rc = LNetMDUnlink(rqbd->rqbd_md_h);
1964 LASSERT (rc == 0 || rc == -ENOENT);
1967 /* Wait for the network to release any buffers it's currently
1970 spin_lock(&service->srv_lock);
1971 rc = service->srv_nrqbd_receiving;
1972 spin_unlock(&service->srv_lock);
1977 /* Network access will complete in finite time but the HUGE
1978 * timeout lets us CWARN for visibility of sluggish NALs */
1979 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
1980 cfs_time_seconds(1), NULL, NULL);
1981 rc = l_wait_event(service->srv_waitq,
1982 service->srv_nrqbd_receiving == 0,
1984 if (rc == -ETIMEDOUT)
1985 CWARN("Service %s waiting for request buffers\n",
1989 /* schedule all outstanding replies to terminate them */
1990 spin_lock(&service->srv_lock);
1991 while (!list_empty(&service->srv_active_replies)) {
1992 struct ptlrpc_reply_state *rs =
1993 list_entry(service->srv_active_replies.next,
1994 struct ptlrpc_reply_state, rs_list);
1995 ptlrpc_schedule_difficult_reply(rs);
1997 spin_unlock(&service->srv_lock);
1999 /* purge the request queue. NB No new replies (rqbds all unlinked)
2000 * and no service threads, so I'm the only thread noodling the
2001 * request queue now */
2002 while (!list_empty(&service->srv_req_in_queue)) {
2003 struct ptlrpc_request *req =
2004 list_entry(service->srv_req_in_queue.next,
2005 struct ptlrpc_request,
2008 list_del(&req->rq_list);
2009 service->srv_n_queued_reqs--;
2010 service->srv_n_active_reqs++;
2011 ptlrpc_server_finish_request(req);
2013 while (ptlrpc_server_request_pending(service, 1)) {
2014 struct ptlrpc_request *req;
2016 req = ptlrpc_server_request_get(service, 1);
2017 list_del(&req->rq_list);
2018 service->srv_n_queued_reqs--;
2019 service->srv_n_active_reqs++;
2020 ptlrpc_hpreq_fini(req);
2021 ptlrpc_server_finish_request(req);
2023 LASSERT(service->srv_n_queued_reqs == 0);
2024 LASSERT(service->srv_n_active_reqs == 0);
2025 LASSERT(service->srv_n_history_rqbds == 0);
2026 LASSERT(list_empty(&service->srv_active_rqbds));
2028 /* Now free all the request buffers since nothing references them
2030 while (!list_empty(&service->srv_idle_rqbds)) {
2031 struct ptlrpc_request_buffer_desc *rqbd =
2032 list_entry(service->srv_idle_rqbds.next,
2033 struct ptlrpc_request_buffer_desc,
2036 ptlrpc_free_rqbd(rqbd);
2039 /* wait for all outstanding replies to complete (they were
2040 * scheduled having been flagged to abort above) */
2041 while (atomic_read(&service->srv_outstanding_replies) != 0) {
2042 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(10), NULL, NULL);
2044 rc = l_wait_event(service->srv_waitq,
2045 !list_empty(&service->srv_reply_queue), &lwi);
2046 LASSERT(rc == 0 || rc == -ETIMEDOUT);
2049 ptlrpc_server_handle_reply(service);
2052 CWARN("Unexpectedly long timeout %p\n", service);
2055 list_for_each_entry_safe(rs, t, &service->srv_free_rs_list, rs_list) {
2056 list_del(&rs->rs_list);
2057 OBD_FREE(rs, service->srv_max_reply_size);
2060 /* In case somebody rearmed this in the meantime */
2061 cfs_timer_disarm(&service->srv_at_timer);
2063 OBD_FREE_PTR(service);
2067 /* Returns 0 if the service is healthy.
2069 * Right now, it just checks to make sure that requests aren't languishing
2070 * in the queue. We'll use this health check to govern whether a node needs
2071 * to be shot, so it's intentionally non-aggressive. */
2072 int ptlrpc_service_health_check(struct ptlrpc_service *svc)
2074 struct ptlrpc_request *request;
2075 struct timeval right_now;
2081 do_gettimeofday(&right_now);
2083 spin_lock(&svc->srv_lock);
2084 if (!ptlrpc_server_request_pending(svc, 1)) {
2085 spin_unlock(&svc->srv_lock);
2089 /* How long has the next entry been waiting? */
2090 if (list_empty(&svc->srv_request_queue))
2091 request = list_entry(svc->srv_request_hpq.next,
2092 struct ptlrpc_request, rq_list);
2094 request = list_entry(svc->srv_request_queue.next,
2095 struct ptlrpc_request, rq_list);
2096 timediff = cfs_timeval_sub(&right_now, &request->rq_arrival_time, NULL);
2097 spin_unlock(&svc->srv_lock);
2099 if ((timediff / ONE_MILLION) > (AT_OFF ? obd_timeout * 3/2 :
2101 CERROR("%s: unhealthy - request has been waiting %lds\n",
2102 svc->srv_name, timediff / ONE_MILLION);